Coverage and limitations

What our synthetic tests cover

ExtractMint's evidence comes from automated tests using invented statement data. Synthetic evidence does not establish real-bank accuracy, real-world OCR accuracy, or a supported-bank certification list.

Evidence checked: 2026-10-28

Digital PDF regression fixtures

The generated English and Malay baseline ledger PDFs contain three invented transactions each: two PDFs and six transaction rows in total. Additional synthetic PDF regressions include a six-transaction US-date fixture, a ten-transaction multiline fixture, and separate 132-transaction completeness fixtures, including a DD/MM date variant.

The automated checks exercise separate debit/credit columns, signed amounts, parentheses negatives, running balances, repeated page headers, multiline continuations, date conventions, and transaction completeness for those fixtures.

These counts describe automated cases rather than customer statements. They are not counts of supported banks, customers, or independently validated real statements.

Synthetic scan exercise

One generated image-only statement runs through the browser OCR path in the sample checks at mobile and desktop widths. The exercise checks detected synthetic content, a visible scan warning, an unresolved date choice, and an unreadable-amount flag.

The result must stay in review. QBO stays blocked while amounts or required date choices remain unresolved, including after the scan warning is acknowledged. These assertions cover content detection and warning/gating behavior for this invented image; they do not measure real-world OCR accuracy across scans, languages, banks, or capture conditions.

In the converter, choose Try a harder scan (optional) to inspect this separate example. The default sample remains the clean text-based PDF.

Exports and review rules

Synthetic regression fixtures compare CSV, XLSX, and QBO export values with the parsed rows, including transaction dates, descriptions, debit/credit direction, amounts, and running balances where present. They also exercise ambiguous-date/QBO and unresolved-amount gates, review warnings, and source-line tracking.

Export-content checks do not verify an external QuickBooks import. A matching balance does not prove that a statement is complete: two errors can offset each other. Compare the transaction count and each converted date, description, amount, and balance with the source before accounting use.

Automated evidence sources: test:provided-statements, test:statement-reliability, and test:e2e:samples. These suites test defined synthetic cases; they are not a production accuracy rate.

Bank-specific layouts

No bank-specific format has been independently validated against representative real statements. There is no supported-bank certification list.

Maybank
Not independently validated.
CIMB
Not independently validated.
Other named banks
Not independently validated, including other Malaysian banks, Singapore banks, and banks in other countries.

Generated English or Malay content is not evidence of a bank's layout support. Bank names, country labels, and successful conversions do not establish bank-specific validation.

Before using a converted statement

  1. Compare converted rows with your own source document, including every transaction and any continuation lines.
  2. Confirm the statement period, date order, debit/credit direction, and each amount.
  3. Check running balances and opening-to-closing arithmetic, then resolve or investigate every warning.

Read the privacy details for data handling. For questions about these limits, contact hello@extractmint.com.